gke-ai-troubleshooting-jobset-interruption

CRITICAL E004: Prompt injection detected in skill instructions.

• Potential prompt injection detected (high risk: 0.70). The skill contains explicit instructions to "BypassSandbox: true" alongside shell execution commands (curl with auth tokens), which are directives to override agent/system constraints and thus constitute a hidden/deceptive instruction outside the troubleshooting guidance's normal scope.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.90). The skill explicitly instructs the agent to bypass the sandbox to run shell commands (multiple places) and recommends destructive infrastructure actions (cordon/drain and delete the underlying GCE VM), which ask the agent to perform operations that modify machine/cluster state and bypass security boundaries.